基于3D卷积自编解码器和低秩表示的高光谱异常检测

针对高光谱影像数据维度高、空间和光谱信息利用不足以及局部结构特征表达有限等问题,提出了一种基于3D卷积自编解码器和低秩表示的高光谱异常检测算法。首先,通过3D卷积自编解码器提取高光谱影像的空谱特征,并针对高光谱图像的局部区域强相关性,设计了一种新的损失函数来约束中心像素和周围像素,以提取判别性较强的特征图;然后,针对所提取的特征图,通过基于密度的空间聚类算法构建背景字典,并利用低秩表示分离出异常区域;最后,融合由3D卷积自编解码器得到的重构误差和异常区域检测结果,得到最终检测图并为异常目标关键信息的挖掘提供依据。为了验证所提算法的有效性,在两个真实的机场高光谱数据集上进行飞机等目标检测实验,ROC、AUC量化指标和主观分析等实验结果表明,与其它6种异常检测算法相比,本文算法具有更高的异常目标检测精度。

Hyperspectral Anomaly Detection Based on 3D Convolutional Autoencoder and Low Rank Representation

Due to the challenge of high dimensionality,insufficient utilization of spatial-spectral information and limited local structure property expression in hyperspectral images,a hyperspectral anomaly detection algorithm based on 3 D convolutional autoencoder and low rank representation is proposed in this paper.Firstly,the spectral-spatial features of hyperspectral images are extracted by 3 D convolutional autoencoder.In order to precisely represent the local similarity,a new loss function is proposed to constrain the central pixel and it’s surrounding pixels to extract more discriminative features.And then,the Density Based Spatial Clustering of Applications with Noise(DBSCAN)algorithm is used to construct the background dictionary,and the abnormal region is separated by low rank representation on the feature map.Finally,the detection result is obtained by fusing the reconstruction error obtained by 3 D convolution autoencoder and abnormal region detection result.We carry out objective and subjective anomaly detection experiments on two real hyperspectral datasets.The results demonstrate that the proposed algorithm detect abnormal targets more accurately compared with other algorithms.